[Paper Review] A Study of Adaptive Modeling Towards Robust Generalization

Zihao Jing, Qiuhao Zeng+5 more|arXiv (Cornell University)|Feb 2, 2026

Machine Learning in Materials Science0 citations

TL;DR

This paper introduces Cuttlefish, a unified all-atom structure-aware LLM that adaptively grounds geometry in language reasoning via Scaling-Aware Patching and a Geometry Grounding Adapter to improve cross-modality grounding and reduce structural hallucinations.

ABSTRACT

Large language models (LLMs) increasingly support reasoning over biomolecular structures, but most existing approaches remain modality-specific and rely on either sequence-style encodings or fixed-length connector tokens for structural inputs. These designs can under-expose explicit geometric cues and impose rigid fusion bottlenecks, leading to over-compression and poor token allocation as structural complexity grows. We present a unified all-atom framework that grounds language reasoning in geometric information while adaptively scaling structural tokens. The method first constructs variable-size structural patches on molecular graphs using an instruction-conditioned gating policy, enabling complexity-aware allocation of query tokens. It then refines the resulting patch tokens via cross-attention with modality embeddings and injects geometry-informed tokens into the language model to improve structure grounding and reduce structural hallucinations. Across diverse all-atom benchmarks, the proposed approach yields consistent gains in heterogeneous structure-grounded reasoning. An anonymized implementation is provided in the supplementary material.

Motivation & Objective

• Motivation to move beyond modality-specific structure inputs toward a unified all-atom interface for LLMs.
• Develop methods that adaptively allocate structural tokens to match molecular complexity.
• Reduce geometric hallucinations by injecting verifiable geometry into language reasoning.
• Demonstrate robust structure-grounded reasoning across multiple all-atom modalities.
• Provide an open all-atom instruction dataset GEO-AT to catalyze future work.

Proposed method

• Introduce Scaling-Aware Patching to allocate query tokens proportional to structural complexity via an instruction-conditioned gating policy.
• Implement a soft patch-growing mechanism to form variable-size structural patches on molecular graphs.
• Develop a Geometry Grounding Adapter that cross-attends to modality embeddings and injects geometry-informed tokens into the LLM.
• Use an SE(3)-equivariant EGNN encoder to produce atom-level modality embeddings.
• Train in two stages: connector-focused end-to-end tuning on GEO-AT, then LLM-adaptation with unfreezing of the LLM for end-to-end optimization.

Figure 1 : Mol-Llama performance on the Mol-Instructions captioning task, evaluated across five molecule length bins with 6 metrics (left y-axis, detailed in App D.3 ) plotted as curves with dashed overall averages, and the background bars indicate the proportion of samples in each length bin (right

Experimental results

Research questions

• RQ1 How can all-atom geometry be effectively represented and fed into LLMs without fixed-length bottlenecks?
• RQ2 Can geometry-grounded tokens reduce structural hallucinations in multimodal molecular reasoning?
• RQ3 Do scaling-aware, patch-based representations improve reasoning across molecules, proteins, DNA, and RNA?
• RQ4 What is the impact of geometry-grounded connectors on performance compared with modality-specific baselines?

Key findings

• Cuttlefish achieves consistent gains in structure-grounded reasoning across all-atom benchmarks and modalities.
• Scaling-Aware Patching mitigates fixed-budget bottlenecks by adaptively allocating tokens to structurally informative regions.
• Geometry Grounding Adapter injects geometry cues into the LLM, reducing hallucinations and improving grounding.
• Cuttlefish shows strong performance gains over general LLMs and modality-specific baselines across molecules, proteins, and nucleic acids.
• The approach maintains stable performance across increasing structural complexity, indicating favorable scaling properties.

Figure 2 : Architecture of Cuttlefish. The framework accepts all-atom inputs (spatial graph: atom features, coordinates, and spatial relations) processed by EGNN for modality embeddings. The model incorporates Scaling-Aware Patching through an instruction-conditioned gate and soft patch-growing mech